The Startups Taking on Scale AI—In Robotics

As I wrote last month, there’s been a flurry of excitement about applying generative artificial intelligence techniques to robotics, a field that’s generated much hope and plenty of disappointment over the last two decades. 

Like chatbot makers, robotics companies have relied on humans to manually label data, for example identifying packages in images to train robots to sort boxes. Many have used Scale AI, the data-labelling startup valued at $14 billion also used by AI model makers like OpenAI.

Now, Scale AI faces growing competition from smaller startups that specialize in labeling data for robotics, including Segments.ai, Kognic and Encord. Warehouse robotics maker Dexterity and self-driving truck company Kodiak Robotics are among the companies that have switched to the startups, which provide software tailored for labeling robotics data. 

Dexterity started using Scale more than four years ago to identify items in scenes, but was not impressed with the quality of the labels, which seemed almost AI-generated, said Robert Sun, a founding engineer at Dexterity. A couple of years later, Dexterity switched to Segments, which farms out much labeling work to smaller companies and offers software that Dexterity’s employees can use to label images themselves.